Universal speech models for speaker independent single channel source separation

Abstract

Supervised and semi-supervised source separation algorithms based on non-negative matrix factorization have been shown to be quite effective. However, they require isolated training examples of one or more sources, which is often difficult to obtain. This limits the practical applicability of these algorithms. We examine the problem of efficiently utilizing general training data in the absence of specific training examples. Specifically, we propose a method to learn a universal speech model from a general corpus of speech and show how to use this model to separate speech from other sound sources. This model is used in lieu of a speech model trained on speaker-dependent training examples, and thus circumvents the aforementioned problem. Our experimental results show that our method achieves nearly the same performance as when speaker-dependent training examples are used. Furthermore, we show that our method improves performance when training data of the non-speech source is available.

Non-Negative Group Sparsity with Subspace Note Modelling for Polyphonic Transcription

“…Other mixed norm penalties, such as the 1,2 term proposed in [21] which seeks to have many groups active with few atoms in each group supported, are known. A list of some of these penalties is given in [22].…”

“…For KL (β = 1), the 2,1 penalty is employed in (29), giving a convex cost function and linear scaling, unlike the approach in [22]. KL with 2,1 penalty was previously used for group sparse NMD [46], however a monotonic algorithm was not developed in [46], and is offered here.…”

“…A true positive is denoted when a detected onset falls within a tolerance of 50 ms from a ground truth onset, with other detections denoted as false positives, and undetected ground truth onsets denoted as false negatives. Analysis of onset detection is performed in a similar manner to the framewise case, using thresholding relative to a range of values of the δ parameter (22), and with results given for P, R, F.…”

“…The optimal results in terms of F-measure, found at δ opt applied across all pieces, were recorded. For GBF-NNLS, with both dictionaries, the stopping condition is calculated from the coefficient matrix of the NNLS decomposition (22), with experiments similarly run for various values of δ. Again the STFT spectrograms were employed.…”

“…A strategy for estimation of the parameters λ, a, in (26) is given in [43], however it is unclear whether this strategy extends to cost functions other than IS. The penalty (26) is also employed with KL divergence in [22], in which case optimisation is performed by using a convex-concave projection algorithm. As an alternative, we now propose a group penalty using an 2 norm group coefficient that is scale invariant to β-divergence.…”